Chen et al. introduce PhysLoop, a unified framework enabling recursive self-improvement (RSI) in video generation models by harnessing large language models' (LLMs) physical knowledge to address persistent failures in physical dynamics and causal consistency. The system constrains LLMs across three stages—generation, selection, and correction—to distill physical reasoning into training data without extensive human supervision. PhysLoop constructs controllable training datasets and supports inference-time localized video repair by identifying physical inconsistencies and synthesizing corrected variants. The authors release 100,000 generated videos covering diverse indoor and outdoor scenes, alongside generation tooling and two benchmarks: PhysLoop-Gen for evaluating controllable generation and PhysLoop-Repair for assessing targeted correction capabilities. Evaluation on these datasets demonstrates improved physical plausibility in generated dynamics, while results on the external Physics-IQ Verified benchmark confirm enhanced adherence to physical laws. By closing the loop between LLM-based physical reasoning and video synthesis, PhysLoop establishes a scalable pipeline for self-improving video models that reduces reliance on manual annotation.